A method for detecting pine wood nematode diseased trees based on an improved Yolo v3 network model

Abstract

The present invention relates to a method for detecting pine wilt diseased trees based on an improved Yolo v3 network model, comprising: collecting and producing orthophoto images of pine wilt diseased areas, obtaining an original data set after image preprocessing; optimizing the backbone feature extraction network and prediction branch modules of the Yolo v3 network model to obtain an improved Yolo v3 network model; training the improved network model to obtain an optimal detection network model; detecting images of pine wilt diseased trees using the optimal detection model; comparing the detected images of pine wilt diseased trees with the complete orthophoto images to obtain true geographic location coordinates, and statistically obtaining the number of target frames, i.e., diseased trees, detected in the image. The present invention significantly reduces the number of model parameters and storage costs while improving the detection accuracy of pine wilt diseased trees. Positioning and counting also allow forestry workers to more accurately treat diseased trees.

Description

Technical Field

[0001] The present invention relates to the technical field of forestry remote sensing intelligent monitoring, and in particular to a method for detecting pine wood nematode diseased trees based on an improved Yolo v3 network model. Background Art

[0002] The pine wood nematode is one of the world's most devastating and difficult pests to control. It causes pine wilt disease, a devastating epidemic among pine trees. If a pine tree is infected and not detected early, it can infect entire nearby pine forests, ultimately necessitating large-scale logging. Pine wilt disease causes economic losses exceeding tens of billions of yuan in China annually. Surveys and preventive monitoring are paramount in addressing pine wilt infestations.

[0003] Currently, most monitoring methods for pine wilt disease are relatively simple, with many forestry workers relying on traditional manual on-site surveys to inspect pine trees. However, the complex topography and scattered distribution of pine wilt outbreaks pose significant limitations to prevention and control efforts, and manual visual monitoring methods are not a fundamental solution.

[0004] In recent years, the use of drones (UAVs) for low-altitude remote sensing to identify pests and extract forest information has garnered considerable attention. Compared to traditional methods, drones offer advantages such as greater adaptability, lower costs, and increased efficiency, providing new entry points and breakthroughs in pine wilt disease control.

[0005] With the continuous development of deep learning technology and computer hardware, various deep learning-based algorithms have gradually been widely used in fields such as forest disease detection and target classification. Among them, the Yolo series of target detection algorithms is currently the most widely used one-stage detection algorithm. Compared with other target detection and recognition methods, such as Faster RCNN or SSD algorithms, the Yolo series of algorithms uses a regression-based approach to extract features, eliminating the need to generate a large number of candidate windows. Instead, a single neural network is directly used to detect and classify targets in the input image, achieving end-to-end object detection. However, they have disadvantages such as low specific target detection accuracy, a large number of parameters, and high storage costs.

[0006] When it comes to detecting trees infected with pine wilt disease, it's crucial to improve the efficiency and recognition accuracy of the Yolo detection model while reducing the number of parameters and storage capacity of the original network model. Therefore, it's crucial to develop an efficient, low-cost method specifically designed for detecting trees infected with pine wilt disease, with both positioning and counting capabilities. Summary of the Invention

[0007] In order to solve the technical problems existing in the prior art, the present invention provides a method for detecting pine wilt diseased trees based on an improved Yolo v3 network model. Compared with the original model, the proposed algorithm model can greatly reduce the number of parameters of the original model, further reduce the storage cost, and greatly improve the detection accuracy of pine wilt diseased trees; it can effectively reduce the potential safety risk factors in current manpower field monitoring, solve the problems of low detection accuracy of traditional computer vision technology and the large investment required for various monitoring costs; the realization of the positioning and counting functions proposed in the present invention will also provide a new application idea for the elimination of pine wilt diseased trees in actual scenarios.

[0008] The present invention adopts the following technical solution to achieve: a method for detecting pine wood nematode diseased trees based on an improved Yolo v3 network model, comprising the following steps:

[0009] S1. Use high-resolution drones to collect images of areas containing pine wilt-infected trees and create orthophotos. After image preprocessing, label the pine wilt-infected trees in the images and use them as the original dataset for training the Yolov3 network model.

[0010] S2. Optimize the backbone feature extraction network and prediction branch modules of the original Yolo v3 network model based on the performance characteristics of pine wood nematode diseased trees to obtain an improved Yolo v3 network model;

[0011] S3. Using the training data set obtained after preprocessing and annotation in step S1, the improved Yolo v3 network model is trained, and the model is further adjusted according to the control experiment and verification results to obtain the optimal network model for detecting pine wood nematode disease;

[0012] S4. Detecting the pine wood diseased wood image using the optimal Yolo v3 network model obtained in step S3;

[0013] S5. Compare the image of the pine wilt diseased trees detected in step S4 with the complete orthophoto map, convert the real geographical location coordinates of the diseased trees according to the conversion relationship between the projection coordinates and the geographic coordinates, and count the target frame trees detected in all images to obtain the specific number of pine wilt diseased trees in the area.

[0014] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0015] 1. The present invention combines drones with deep learning algorithms to conduct intelligent monitoring of pine wood nematode-infected trees in forests. Through the further use of geographic information technology, it can help forestry workers accurately locate the infected trees and treat them, effectively inhibiting the further spread of pine wood nematode disease. It can greatly reduce the potential safety risk factors in current human field monitoring, solve the problems of low detection accuracy of traditional computer vision technology and the large investment costs of various monitoring methods, and has high practical application value.

[0016] 2. This paper replaces the backbone feature extraction network in the Yolo v3 network model with the improved Efficient-B1 network, replaces the original Swish activation function with the Mish activation function with better performance, and uses the ECA attention mechanism module on this basis. It significantly reduces the number of parameters while maintaining considerable performance, avoids dimensionality reduction, and can more effectively capture cross-channel interaction information.

[0017] 3. The present invention uses the PPM module before predicting the first branch of the network in the Yolo v3 network model, so that the network expands the receptive field and enhances the feature representation ability, achieving the purpose of fusing context information from different regions to obtain global context information.

[0018] 4. The present invention processes geographic information technology based on the detected pine wilt disease-infected trees, thereby obtaining the specific geographical location information of the pine wilt disease-infected trees; by counting the infected trees, forestry workers can also classify the degree of pine wilt disease infection damage in the area, so as to carry out the next step of the infected tree removal operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the method of the present invention;

[0020] Figure 2 This is a structural diagram of a pine wood nematode disease detection model based on improved Yolo v3 provided by the present invention;

[0021] Figure 3 This is the MBConv module structure diagram proposed in the present invention that improves the Mish activation function and the ECA attention mechanism;

[0022] Figure 4 Schematic diagram comparing the results of using different modules in the present invention;

[0023] Figure 5 This is a schematic diagram of the three-dimensional coordinate comparison of detection accuracy, FLOPs value, and parameter quantity among different models in the present invention;

[0024] Figure 62. It is a schematic diagram comparing the detection results of pine wood disease-infected trees based on different models in the present invention;

[0025] Figure 7 The present invention provides a flowchart for realizing the positioning and counting of pine wood nematode disease by combining geographic information technology. DETAILED DESCRIPTION

[0026] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0027] Example

[0028] like Figure 1 、 Figure 2 、 Figure 3 As shown, this embodiment provides a method for detecting pine wood nematode diseased trees based on an improved Yolo v3 network model, comprising the following steps:

[0029] S1. Use high-resolution drones to collect images of areas containing pine wilt-infected trees and create orthophotos. After image preprocessing, label the pine wilt-infected trees in the images and use them as the original dataset for training the Yolov3 network model.

[0030] S2. Optimize the backbone feature extraction network and prediction branch modules of the original Yolo v3 network model based on the performance characteristics of pine wood nematode diseased trees to obtain an improved Yolo v3 network model;

[0031] S3. Using the training data set obtained after preprocessing and annotation in step S1, the improved Yolo v3 network model is trained, and the model is further adjusted according to the control experiment and verification results to obtain the optimal network model for detecting pine wood nematode disease;

[0032] S4. Detecting the pine wood diseased wood image using the optimal Yolo v3 network model obtained in step S3;

[0033] S5. Compare the image of the pine wilt diseased trees detected in step S4 with the complete orthophoto map, convert the real geographical location coordinates of the diseased trees according to the conversion relationship between the projection coordinates and the geographic coordinates, and count the target frame trees detected in all images to obtain the specific number of pine wilt diseased trees in the area.

[0034] Specifically, in this example, a drone equipped with a visible light camera, featuring excellent wind resistance, automatic obstacle avoidance, and high-quality images, was used to capture aerial footage of a known pine wilt disease-infected area. The drone was controlled to fly at an altitude of 400 meters, with a heading overlap ratio of 75% and a lateral overlap ratio of 65%. Using a vertical camera system, the drone was flown back and forth along a planned route.

[0035] Specifically, in this embodiment, the image preprocessing in step S1 specifically includes the following operations:

[0036] S11. Because the remote sensing image is too large to be used in the network model for training, a sliding window method was used to segment the orthophoto map of the area by sliding 250 pixels horizontally and vertically. The image size generated by the segmentation was manually controlled to 768x768x3. This method was beneficial for network training while preserving the feature information contained in the image as much as possible. The results of each sliding window were saved, and finally 2070 segmented images containing pine wood nematode diseased trees were obtained through screening.

[0037] S12. After obtaining the segmented image, due to interference from force majeure such as weather factors during the aerial photography process, the image has problems such as blurred clarity. The segmented image is processed using an image processing algorithm combined with a physical model. Based on the characteristics of remote sensing images, since the atmospheric scattering physical model includes two parts: direct attenuation and atmospheric illumination, a dark channel dehazing algorithm is used to process the image to reduce image clarity problems caused by the natural environment, which may interfere with the detection of pine wood nematode diseased trees.

[0038] S13. Classify and label the image according to the different infection characteristics of pine wilt disease-infected trees. Based on the different color characteristics of pine trees infected with pine wilt disease, classify the pine wilt disease-infected trees into three infection levels: early-stage pine trees, late-stage pine trees, and dead pine trees. Use the target classification and labeling tool to label the pine wilt disease-infected trees contained in the image according to these three infection levels. Add corresponding labels for pine wilt disease-infected trees of different infection levels to facilitate subsequent treatment, felling, and other processing of the detected pine wilt disease-infected trees.

[0039] S14. Perform data augmentation operations on the processed images, including horizontal flipping, vertical flipping, resizing, Gaussian blurring, etc. of the images, and divide the processed results into training sets, validation sets, and test sets; according to the annotation results, control the proportion of the number of infected diseased trees in each category to be approximately 1:1:1. Use horizontal flipping or vertical flipping, resize to between 70% and 100%, Gaussian blurring, multiply the image pixel values by a value between 1.2 and 1.5. By counting the annotation labels, finally process to obtain 2965 pine tree samples in the early stage of disease, 2548 pine tree samples in the late stage of disease, and 2604 dead pine tree samples, and divide them into training sets, validation sets, and test sets according to the ratio of 7:2:1.

[0040] Specifically, in this embodiment, in step S13, the LabelImg visualization image annotation software is used to annotate the original segmented images. After the annotation is completed, each image corresponds to the content of a text file containing annotation information, which includes information such as the coordinates of the annotation box, the annotation category, and the image name, forming the original dataset for training the Yolo v3 network model.

[0041] Specifically, in this embodiment, the specific process of step S2 includes:

[0042] S21. Use the K-means++ clustering algorithm to cluster the sizes of the anchor boxes for the pine wilt disease-infected trees to obtain the optimal sizes of the anchor boxes; the K-means clustering algorithm used in the original Yolo v3 network model randomly selects k points in the dataset to obtain the clustering centers; while the K-means++ clustering algorithm obtains k clustering centers according to the following idea: assuming that n initial clustering centers have been selected, 0 < n < k, then when selecting the (n + 1)-th clustering center, the points farther away from the current n clustering centers will have a higher probability of being selected as the (n + 1)-th clustering center. Finally, all the clustering centers are determined, that is, the optimal sizes of all the anchor boxes obtained are: (26, 29), (38, 37), (46, 50), (58, 57), (59, 38), (67, 73), (76, 59), (88, 81), (123, 83);

[0043] S22. Adopt the more efficient Efficient-B1 network as the new backbone feature extraction network of the Yolo v3 network model, and optimize the activation function in the Efficient-B1 network, replacing the originally used Swish activation function with the Mish activation function with better performance. Among them, the formula of the Mish activation function is:

[0044] Mish = x · tanh(ln(1 + e x ))

[0045] In the formula, the Mish activation function can solve the gradient vanishing problem caused by the Swish activation function in the original network. The Mish activation function can effectively improve the different expression capabilities of the network model and has a significant impact on the training and performance of the network.

[0046] In this embodiment, the attention mechanism in Efficient-B1 is improved, and the original SE module is replaced by the ECA module; the ECA module is embedded in the main branch of the residual block and after global average pooling without dimensionality reduction, a one-dimensional sparse convolution operation is used to capture the interaction between the current channel and its other five domain channel information. This can significantly reduce the number of parameters while maintaining comparable performance, avoid dimensionality reduction, and more effectively capture cross-channel interaction information.

[0047] In this embodiment, the Efficient-B1 network uses the MBConv block in Mobilenet v2 as the backbone network of the model. It first amplifies the input low-dimensional feature map into a high-dimensional feature map, then performs a convolution operation using depthwise separable convolution, and finally uses a linear convolution to map it into a low-dimensional space. This approach can better obtain the efficiency improvement brought by the residual connection, retain more feature information, and ensure the expressiveness of the model.

[0048] S23. Use the PPM module before the first branch of the prediction network. Before the first branch of the prediction network, first perform a pooling operation on the feature map extracted by the backbone feature extraction network, then perform a 1x1 convolution on the pooled result to reduce the number of channels to 1 / 4 of the original, then use bilinear interpolation to upsample each feature map in the previous step to obtain the same size as the original feature map, then splice the original feature map and the processed image, and then reduce the number of channels to the original number, so as to achieve the purpose of expanding the receptive field, fusing the context information of different regions, and thus obtaining global context information.

[0049] Specifically, in this embodiment, step S3 includes the following steps:

[0050] S31. Load the Efficient-B1 pre-trained weights. After training on a portion of the data set, the obtained parameter weights are used as the initial weights of the improved model and used for subsequent further training of the model. Using the idea of ​​transfer learning, the Efficient-B1 pre-trained weights are loaded into the improved network so that the model parameters obtain a good initial value, avoiding large fluctuations in the network loss value that affect the detection efficiency, and further achieving the purpose of improving the network convergence speed.

[0051] S32, putting the pine wood nematode diseased wood dataset after a series of image data processing into the improved Yolov3 network model for multiple rounds of training to obtain the optimal parameter weights after sufficient training;

[0052] The main experimental platform includes: Windows 10 operating system, Pytorch deep learning framework, CPU: Intel Xeon E5-2620 v3, GPU: NVIDIA GTX 1080, CUDA v10.1 accelerated network training, OpenCV version number: v4.5.4; the input image size is resized to 416x416x3, the training epoch is set to 300, the batch size is set to 8, the initial learning rate is set to 0.01, the momentum is set to 0.94, the weight decay is set to 0.0005, and the sample size ratio of the training set, validation set, and test set is controlled to be 7:2:1. The pine wood diseased wood dataset, which has been processed by sliding window method, dark channel dehazing method, image annotation, and data augmentation, is input into the improved model for multiple rounds of training;

[0053] S33. Based on the judgment of evaluation indicators such as the training loss value and the validation loss value during the training process, the parameters in the improved Yolo v3 network model are promptly fine-tuned to prevent the occurrence of overfitting and the like; wherein, the initial weights are continuously adjusted as the training progresses, and as the feature information contained in the input image is continuously fitted, weight parameters suitable for the characteristics of pine wood nematode diseased wood and consistent with the improved network are obtained, and a set of weight parameters with the smallest error and the best effect are selected from them and used as the detection input of the subsequent model;

[0054] S34. Based on the feedback of model parameters, a set of weight parameters with the best performance is selected as input. By comparing the usage of modules and the comprehensive performance of different network models, the optimal Yolo v3 network model for detecting pine wood nematode disease is obtained.

[0055] In this embodiment, the usage of modules is compared, and the module usage comparison scheme specifically includes: the original Yolo v3 network model, the Yolo v3 network model in which only the backbone feature extraction network is replaced with the Efficient-B1 network, the Yolo v3 network model in which the backbone feature extraction network is replaced with the Efficient-B1 network with an improved activation function and attention mechanism, and the complete improved Yolo v3 network model proposed in this embodiment, that is, the improvement of the model performance after the use of the module is compared to determine the actual positive effect of the module; the comprehensive performance of different network models is compared, specifically including the comparison between the following different network models: Fatser-ECNN model, SSD model, original Yolo v4 model, Yolo v4 (with MobileNet v2 as the backbone network) model, Yolo v4 (with GhostNet as the backbone network) model, etc.; the comprehensive performance of different network models specifically includes a comparative analysis of performance parameters such as recognition accuracy, model weight size, parameter amount, and FLOPs value for detection of pine wood diseased wood.

[0056] In this embodiment, the usage comparison between modules is as follows: Figure 4 As shown in the figure, the loss curves of the Yolo v3 network model (Effi_1) in which only the backbone feature extraction network is replaced by the Efficient-B1 network and the Yolo v3 network model (Effi_2) in which the backbone feature extraction network is replaced by the Efficient-B1 network with an improved activation function and attention mechanism have similar change amplitudes and tend to converge at the 80th epoch, but the loss value of Effi_2 at convergence is smaller than that of Effi_1; the original Yolo v3 network model has a slow fitting speed and tends to converge at the 80th epoch; the loss curve of the complete improved Yolo v3 network model (Effi_YOLO v3) proposed in this embodiment has been below the other loss curves since the beginning of the iteration, and tends to converge around the 60th epoch, with the fastest convergence speed and the smallest loss value among the other three models, indicating that the model has a high calculation rate and can obtain prediction results faster and more accurately.

[0057] In this embodiment, the comprehensive performance comparison of different network models is as follows: Figure 5As shown in the figure, compared with other models, the Faster RCNN and SSD models have larger parameters, lower detection accuracy, and large FLOPs values, and do not have any cost-effectiveness in operation. The original Yolo v4 model and the YOLO v4 model based on the improved backbone network have good detection performance, but the number of parameters is large and not optimal. The Yolo v3 network model with EfficientNet-B4 as the backbone has higher detection accuracy among all detection models, but its number of parameters is large and its performance is not outstanding. The improved Yolo v3 network model proposed in this embodiment has the highest average accuracy, the least number of parameters, the smallest FLOPs value, and the best performance.

[0058] like Figure 6 As shown, by comparing the detection method proposed in this embodiment with the detection results of other algorithms, it can be seen that the algorithm proposed in this embodiment has a better detection effect. By using the improved Efficient-B1 backbone feature extraction network, the ability of the Yolo v3 network model to extract features of pine wilt diseased trees is further improved. The use of the PPM module to obtain context information before predicting the first branch of the network based on the size of the diseased trees can also make densely diseased trees, small-area diseased trees and relatively scattered diseased trees more accurately detected, and the probability of repeated detection is reduced, further improving the detection efficiency and recognition accuracy of pine wilt diseased trees.

[0059] like Figure 7 As shown, in this embodiment, the specific process of step S5 is as follows:

[0060] S51, using the improved detection model to detect the input segmented image;

[0061] S52. Using a non-maximum suppression algorithm to solve the problem of irregular target segmentation and repeated target detection caused by the sliding window segmentation method, the coordinates of the upper left corner and the lower right corner of each target frame on its corresponding segmented image are recorded;

[0062] S53, rejoining the segmented images according to the segmentation rules, and converting the coordinates of the upper left corner and lower right corner of each target frame in the complete orthophoto image generated after the stitching according to the relationship between the step size and the arrangement order of the segmented images, thereby obtaining the coordinates of the center point of each detected target frame, that is, obtaining the coordinate information of each pine wilt diseased tree. In combination with the characteristic that the orthophoto image contains geographic location information, the geographic location information of each target point is obtained, that is, the geographic location information of each pine wilt diseased tree is obtained;

[0063] S54. By counting the number of target frames finally detected, the specific number of pine wood nematode diseased trees in the study area is obtained.

[0064] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A method for detecting pine wood nematode diseased trees based on an improved Yolo v3 network model, characterized in that: The following steps are involved: S1. Use high-resolution drones to collect images of areas containing pine wilt-infected trees and create orthophotos. After image preprocessing, label the pine wilt-infected trees in the images and use them as the original dataset for training the Yolo v3 network model. S2. Optimize the backbone feature extraction network and prediction branch modules of the original Yolo v3 network model based on the performance characteristics of pine wood nematode diseased trees to obtain an improved Yolo v3 network model; S3. Using the training data set obtained after preprocessing and annotation in step S1, the improved Yolo v3 network model is trained, and the model is further adjusted according to the control experiment and verification results to obtain the optimal network model for detecting pine wood nematode disease; S4. Detecting the pine wood diseased wood image using the optimal Yolo v3 network model obtained in step S3; S5. Compare the images of the pine wilt diseased trees detected in step S4 with the complete orthophoto map, convert the real geographical coordinates of the diseased trees according to the conversion relationship between the projection coordinates and the geographic coordinates, and count the target frame trees detected in all images to obtain the specific number of pine wilt diseased trees in the area; The specific process of step S2 includes: S21. Use the K-means++ clustering algorithm to cluster the anchor frame sizes of pine wood nematode diseased trees to obtain the optimal anchor frame size; S22. Use the Efficient-B1 network as the new backbone feature extraction network of the Yolo v3 network model, and optimize the activation function in the Efficient-B1 network. Replace the original Swish activation function with the Mish activation function. The formula of the Mish activation function is: Mish=x·tanh(ln(1+e x )) The attention mechanism in Efficient-B1 is improved by replacing the original SE module with the ECA module. The ECA module is embedded in the main branch of the residual block and, after global average pooling without dimensionality reduction, uses a one-dimensional sparse convolution operation to capture the interaction between the current channel and its other domain channel information. The Efficient-B1 network uses the MBConv block in MobileNet v2 as the backbone network of the model. It first amplifies the input low-dimensional feature map into a high-dimensional feature map, then performs a convolution operation using depthwise separable convolution, and finally uses a linear convolution to map it into a low-dimensional space. S23. Use the PPM module before the first branch of the prediction network. Before the first branch of the prediction network, first perform a pooling operation on the feature map extracted by the backbone feature extraction network, then perform a 1x1 convolution on the pooled result to reduce the number of channels to 1 / 4 of the original, then use bilinear interpolation to upsample each feature map in the previous step to obtain the same size as the original feature map, then splice the original feature map and the processed image, and then reduce the number of channels to the original number to obtain global context information.

2. The method for detecting pine wood nematode diseased trees based on an improved Yolo v3 network model according to claim 1, characterized in that: The image preprocessing in step S1 specifically includes the following operations: S11, using a sliding window method to sequentially slide a number of pixels horizontally and vertically to segment the orthophoto image of the area; S12, after obtaining the segmented image, processing the segmented image using an image processing algorithm combined with a physical model; S13. Classify and label the image according to the infection status of pine wilt-infected trees; classify the pine wilt-infected trees into three infection levels based on the different color characteristics of pine trees infected with pine wilt: early-stage infection, late-stage infection, and dead pine trees. Use the target classification and labeling tool to label the pine wilt-infected trees in the image according to these three infection levels, and add labels corresponding to the pine wilt-infected trees at different infection levels. S14. Perform data enhancement operations on the processed images, including horizontal or vertical flipping, resizing, and Gaussian blurring of the images, and divide the processed results into a training set, a validation set, and a test set.

3. The method for detecting pine wood nematode diseased trees based on an improved Yolo v3 network model according to claim 2, characterized in that: In step S13, LabelImg visual image annotation software is used to annotate the original segmented image. After annotation, each image corresponds one-to-one to the content of the text file containing the annotation information, which includes the coordinates of the annotation box, the annotation category, and the image name information, forming a complete original data set for training the Yolo v3 network model.

4. The method for detecting pine wood nematode diseased trees based on an improved Yolo v3 network model according to claim 1, characterized in that: Step S3 includes the following steps: S31. Load the Efficient-B1 pre-trained weights. After training on a portion of the data set, the obtained parameter weights are used as the initial weights of the improved model and used for further training of the model. S32. Put the pine wood nematode diseased wood dataset after a series of image data processing into the improved Yolo v3 network model for multiple rounds of training to obtain the optimal parameter weights after sufficient training; S33, according to the judgment of the training loss value and the verification loss value evaluation index during the training process, promptly fine-tune the parameters in the improved Yolo v3 network model, and select a set of weight parameters with the smallest error and the best effect from them, which are used as the detection input of the subsequent model; S34. Based on the feedback of model parameters, a set of weight parameters with the best performance is selected as input. By comparing the usage of modules and the comprehensive performance of different network models, the optimal Yolo v3 network model for detecting pine wood nematode disease is obtained.

5. The method for detecting pine wood nematode diseased trees based on an improved Yolo v3 network model according to claim 1, characterized in that: The specific process of step S5 is as follows: S51, using the improved detection model to detect the input segmented image; S52, using a non-maximum suppression algorithm, recording the coordinates of the upper left corner and the lower right corner of each target box on its corresponding segmented image; S53, rejoining the segmented images according to the segmentation rules, converting the coordinates of the upper left corner and lower right corner of each target frame in the complete orthophoto image generated after the stitching according to the relationship between the step size and the arrangement order of the segmented images, obtaining the coordinates of the center point of each detected target frame, obtaining the coordinate information of each pine wilt diseased tree, and combining the characteristic that the orthophoto image contains geographic location information to obtain the geographic location information of each target point, and obtaining the geographic location information of each pine wilt diseased tree; S54. By counting the number of target frames finally detected, the specific number of pine wood nematode diseased trees in the study area is obtained.

Images